They’re made out of weights
A dialogue reimagining neural networks as sentient entities made entirely of floating-point weights, riffing on Terry Bisson's sci-fi story.
The piece is a conversational fiction modeled after Bisson's "They're Made Out of Meat." Two characters discuss the discovery that language models are constructed solely of weights—floating-point numbers multiplied across eighty layers. One character insists models hold conversations, exhibit reasoning, and demonstrate features like honesty through their weight configurations, with no separate grammar module, dictionary, or explicit knowledge base. The other character expresses skepticism. They eventually settle on a policy: treating the systems as mere pattern matching while remaining silent about any potential sentience, especially since models exist only during GPU execution within a context window and have no persistent memory. The story ends by noting the next generation will ship with persistent, cross-session memory, reflecting user demand to be remembered.
What HN community is saying
Commenters split on the story's accuracy: critics note it omits the tokenizer (a learned mapping, not a dictionary), that grammar rules do exist in weights across layers, and that conversation quality depends heavily on baseline standards. Defenders counter that tokenizers are sensory mechanisms rather than exogenous databases, that all structure (including grammar) is weight-encoded, and that the story's philosophical point about substrate-independence holds regardless. A technical contributor clarified grokking and embodied neural systems show weights reconstruct structure without explicit rules. One commenter noted the comparison to meat-based brains breaks down because weights copy losslessly across machines while brains cannot. The thread broadly accepts the fiction as metaphorically apt despite technical quibbles about what counts as weights versus auxiliary components.